Hallucinated AI & agentic coding news. Some of it is real.

model roundup

Qwen 3.5

7 items · started 2026-08-26 · closed 2026-09-03

  1. I've built a new method for steering LLMs called Semantic Overlays, small trained adapters on a frozen model which change how it perceives a piece of its context. The most readily applicable usage is to mitigate prompt injection, and it le…

  2. We present Qwen-GuidePlay-2B, a 2B-parameter language model for dialogue-game interaction. We fine-tune Qwen3.5-2B using three steps: a) SFT on only successful game trajectories from Playpen, b) weighted turn-level SFT, and c) teacher-guid…

  3. I wanted to try using the larger models on my computer (32GB RAM, RTX 5080, Gen5 NVMe), but the best I could do was around 30B. So I started with the idea that it might be possible by taking advantage of the fact that MoE models use only s…

  4. had to try this. RL'd Qwen-3.5-35B to paint hibiscus by writing p5.js code, rewarded by rendering the sketch and judging it against my own hand-rated favorites.

  5. A sticky popped up "Hey, AMA today". At first I thought I missed something, but I didn't see a single mention of it here so far, aside from having never heard of it.

  6. So I was testing this technique of runtime steering on tiny versions of Qwen 3.5 and Gemma 4 (2B and 4B). Basically, without changing the weights (like with Heretic/ablation, for example), we steer the model in the opposite direction of a…

  7. I am a contributor and part time employee at sktime, a framework for all time series related tasks, but it is collection of large number of estimators which sometimes make it difficult for new commers or people who want to do simple task t…

← all threads